<div dir="ltr"><div>As Amjad noted in <a href="https://reviews.llvm.org/D42739" rel="noreferrer" target="_blank">D42739</a>, I failed to account for select / PHI ops with multi-uses in this patch. I've attempted to close that gap with:<br><a href="https://reviews.llvm.org/rL324252">https://reviews.llvm.org/rL324252</a><br><br></div>Please let me know if the inf-loop is gone with that change. I'm still trying to create a test that would find that gap, but assuming that was the bug, if we can get a reduction from whatever real program hit it, that would be great. Thanks!<br></div><div class="gmail_extra"><br><div class="gmail_quote">On Mon, Feb 5, 2018 at 7:28 AM, Sanjay Patel <span dir="ltr"><<a href="mailto:spatel@rotateright.com" target="_blank">spatel@rotateright.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr">Thanks for letting me know. Sounds like an inf-loop.<br><br>Since this patch is duplicating a transform that occurs in -aggressive-instcombine, it's possible that the bug already exists apart from this change, but might take the benign form of a transform reversal across passes, or it might be masked in that case.<span class="m_5236060557380394234gmail-gI"><span><br></span></span></div><div class="HOEnZb"><div class="h5"><div class="gmail_extra"><br><div class="gmail_quote">On Sat, Feb 3, 2018 at 8:02 PM, Chandler Carruth <span dir="ltr"><<a href="mailto:chandlerc@gmail.com" target="_blank">chandlerc@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr"><div class="gmail_quote"><span><div dir="ltr">On Thu, Feb 1, 2018 at 1:57 PM Sanjay Patel via llvm-commits <<a href="mailto:llvm-commits@lists.llvm.org" target="_blank">llvm-commits@lists.llvm.org</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">Author: spatel<br>
Date: Thu Feb 1 13:55:53 2018<br>
New Revision: 324014<br>
<br>
URL: <a href="http://llvm.org/viewvc/llvm-project?rev=324014&view=rev" rel="noreferrer" target="_blank">http://llvm.org/viewvc/llvm-pr<wbr>oject?rev=324014&view=rev</a><br>
Log:<br>
[InstCombine] allow multi-use values in canEvaluate* if all uses are in 1 inst<br>
<br>
This is the enhancement suggested in D42536 to fix a shortcoming in<br>
regular InstCombine's canEvaluate* functionality.<br>
When we have multiple uses of a value, but they're all in one instruction, we can<br>
allow that expression to be narrowed or widened for the same cost as a single-use<br>
value.<br>
<br>
AFAICT, this can only matter for multiply: sub/and/or/xor/select would be simplified<br>
away if the operands are the same value; add becomes shl; shifts with a variable shift<br>
amount aren't handled.<br></blockquote><div><br></div></span><div>FWIW, we have at least one case which is now either inf-looping or taking dramatically longer to compile after this commit (way over 2x from what I can tell).</div><div><br></div><div>We're working on building a test case, but just wanted to send a heads-up in case others are hitting slow compiles.</div><div><div class="m_5236060557380394234h5"><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<br>
Differential Revision: <a href="https://reviews.llvm.org/D42739" rel="noreferrer" target="_blank">https://reviews.llvm.org/D4273<wbr>9</a><br>
<br>
Modified:<br>
llvm/trunk/lib/Transforms/Inst<wbr>Combine/InstCombineCasts.cpp<br>
llvm/trunk/test/Transforms/Ins<wbr>tCombine/cast-mul-select.ll<br>
llvm/trunk/test/Transforms/Ins<wbr>tCombine/icmp-mul-zext.ll<br>
<br>
Modified: llvm/trunk/lib/Transforms/Inst<wbr>Combine/InstCombineCasts.cpp<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/InstCombine/InstCombineCasts.cpp?rev=324014&r1=324013&r2=324014&view=diff" rel="noreferrer" target="_blank">http://llvm.org/viewvc/llvm-pr<wbr>oject/llvm/trunk/lib/Transform<wbr>s/InstCombine/InstCombineCasts<wbr>.cpp?rev=324014&r1=324013&r2=<wbr>324014&view=diff</a><br>
==============================<wbr>==============================<wbr>==================<br>
--- llvm/trunk/lib/Transforms/Inst<wbr>Combine/InstCombineCasts.cpp (original)<br>
+++ llvm/trunk/lib/Transforms/Inst<wbr>Combine/InstCombineCasts.cpp Thu Feb 1 13:55:53 2018<br>
@@ -185,8 +185,14 @@ Value *InstCombiner::EvaluateInDiffe<wbr>rent<br>
case Instruction::Shl:<br>
case Instruction::UDiv:<br>
case Instruction::URem: {<br>
- Value *LHS = EvaluateInDifferentType(I->get<wbr>Operand(0), Ty, isSigned);<br>
- Value *RHS = EvaluateInDifferentType(I->get<wbr>Operand(1), Ty, isSigned);<br>
+ Value *LHS, *RHS;<br>
+ if (I->getOperand(0) == I->getOperand(1)) {<br>
+ // Don't create an unnecessary value if the operands are repeated.<br>
+ LHS = RHS = EvaluateInDifferentType(I->get<wbr>Operand(0), Ty, isSigned);<br>
+ } else {<br>
+ LHS = EvaluateInDifferentType(I->get<wbr>Operand(0), Ty, isSigned);<br>
+ RHS = EvaluateInDifferentType(I->get<wbr>Operand(1), Ty, isSigned);<br>
+ }<br>
Res = BinaryOperator::Create((Instru<wbr>ction::BinaryOps)Opc, LHS, RHS);<br>
break;<br>
}<br>
@@ -320,10 +326,12 @@ static bool canNotEvaluateInType(Value *<br>
assert(!isa<Constant>(V) && "Constant should already be handled.");<br>
if (!isa<Instruction>(V))<br>
return true;<br>
- // We can't extend or shrink something that has multiple uses: doing so would<br>
- // require duplicating the instruction in general, which isn't profitable.<br>
+ // We can't extend or shrink something that has multiple uses -- unless those<br>
+ // multiple uses are all in the same instruction -- doing so would require<br>
+ // duplicating the instruction which isn't profitable.<br>
if (!V->hasOneUse())<br>
- return true;<br>
+ if (any_of(V->users(), [&](User *U) { return U != V->user_back(); }))<br>
+ return true;<br>
<br>
return false;<br>
}<br>
<br>
Modified: llvm/trunk/test/Transforms/Ins<wbr>tCombine/cast-mul-select.ll<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/InstCombine/cast-mul-select.ll?rev=324014&r1=324013&r2=324014&view=diff" rel="noreferrer" target="_blank">http://llvm.org/viewvc/llvm-pr<wbr>oject/llvm/trunk/test/Transfor<wbr>ms/InstCombine/cast-mul-<wbr>select.ll?rev=324014&r1=324013<wbr>&r2=324014&view=diff</a><br>
==============================<wbr>==============================<wbr>==================<br>
--- llvm/trunk/test/Transforms/Ins<wbr>tCombine/cast-mul-select.ll (original)<br>
+++ llvm/trunk/test/Transforms/Ins<wbr>tCombine/cast-mul-select.ll Thu Feb 1 13:55:53 2018<br>
@@ -49,11 +49,9 @@ define i8 @select2(i1 %cond, i8 %x, i8 %<br>
<br>
define i32 @eval_trunc_multi_use_in_one_i<wbr>nst(i32 %x) {<br>
; CHECK-LABEL: @eval_trunc_multi_use_in_one_i<wbr>nst(<br>
-; CHECK-NEXT: [[Z:%.*]] = zext i32 [[X:%.*]] to i64<br>
-; CHECK-NEXT: [[A:%.*]] = add nuw nsw i64 [[Z]], 15<br>
-; CHECK-NEXT: [[M:%.*]] = mul i64 [[A]], [[A]]<br>
-; CHECK-NEXT: [[T:%.*]] = trunc i64 [[M]] to i32<br>
-; CHECK-NEXT: ret i32 [[T]]<br>
+; CHECK-NEXT: [[A:%.*]] = add i32 [[X:%.*]], 15<br>
+; CHECK-NEXT: [[M:%.*]] = mul i32 [[A]], [[A]]<br>
+; CHECK-NEXT: ret i32 [[M]]<br>
;<br>
%z = zext i32 %x to i64<br>
%a = add nsw nuw i64 %z, 15<br>
@@ -64,11 +62,9 @@ define i32 @eval_trunc_multi_use_in_one_<br>
<br>
define i32 @eval_zext_multi_use_in_one_in<wbr>st(i32 %x) {<br>
; CHECK-LABEL: @eval_zext_multi_use_in_one_in<wbr>st(<br>
-; CHECK-NEXT: [[T:%.*]] = trunc i32 [[X:%.*]] to i16<br>
-; CHECK-NEXT: [[A:%.*]] = and i16 [[T]], 5<br>
-; CHECK-NEXT: [[M:%.*]] = mul nuw nsw i16 [[A]], [[A]]<br>
-; CHECK-NEXT: [[R:%.*]] = zext i16 [[M]] to i32<br>
-; CHECK-NEXT: ret i32 [[R]]<br>
+; CHECK-NEXT: [[A:%.*]] = and i32 [[X:%.*]], 5<br>
+; CHECK-NEXT: [[M:%.*]] = mul nuw nsw i32 [[A]], [[A]]<br>
+; CHECK-NEXT: ret i32 [[M]]<br>
;<br>
%t = trunc i32 %x to i16<br>
%a = and i16 %t, 5<br>
@@ -79,12 +75,10 @@ define i32 @eval_zext_multi_use_in_one_i<br>
<br>
define i32 @eval_sext_multi_use_in_one_in<wbr>st(i32 %x) {<br>
; CHECK-LABEL: @eval_sext_multi_use_in_one_in<wbr>st(<br>
-; CHECK-NEXT: [[T:%.*]] = trunc i32 [[X:%.*]] to i16<br>
-; CHECK-NEXT: [[A:%.*]] = and i16 [[T]], 14<br>
-; CHECK-NEXT: [[M:%.*]] = mul nuw nsw i16 [[A]], [[A]]<br>
-; CHECK-NEXT: [[O:%.*]] = or i16 [[M]], -32768<br>
-; CHECK-NEXT: [[R:%.*]] = sext i16 [[O]] to i32<br>
-; CHECK-NEXT: ret i32 [[R]]<br>
+; CHECK-NEXT: [[A:%.*]] = and i32 [[X:%.*]], 14<br>
+; CHECK-NEXT: [[M:%.*]] = mul nuw nsw i32 [[A]], [[A]]<br>
+; CHECK-NEXT: [[O:%.*]] = or i32 [[M]], -32768<br>
+; CHECK-NEXT: ret i32 [[O]]<br>
;<br>
%t = trunc i32 %x to i16<br>
%a = and i16 %t, 14<br>
<br>
Modified: llvm/trunk/test/Transforms/Ins<wbr>tCombine/icmp-mul-zext.ll<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/InstCombine/icmp-mul-zext.ll?rev=324014&r1=324013&r2=324014&view=diff" rel="noreferrer" target="_blank">http://llvm.org/viewvc/llvm-pr<wbr>oject/llvm/trunk/test/Transfor<wbr>ms/InstCombine/icmp-mul-zext.<wbr>ll?rev=324014&r1=324013&r2=<wbr>324014&view=diff</a><br>
==============================<wbr>==============================<wbr>==================<br>
--- llvm/trunk/test/Transforms/Ins<wbr>tCombine/icmp-mul-zext.ll (original)<br>
+++ llvm/trunk/test/Transforms/Ins<wbr>tCombine/icmp-mul-zext.ll Thu Feb 1 13:55:53 2018<br>
@@ -55,13 +55,12 @@ lor.end:<br>
<br>
define void @PR33765(i8 %beth) {<br>
; CHECK-LABEL: @PR33765(<br>
-; CHECK-NEXT: [[CONV:%.*]] = zext i8 [[BETH:%.*]] to i32<br>
+; CHECK-NEXT: [[CONV:%.*]] = zext i8 [[BETH:%.*]] to i16<br>
; CHECK-NEXT: br i1 false, label [[IF_THEN9:%.*]], label [[IF_THEN9]]<br>
; CHECK: if.then9:<br>
-; CHECK-NEXT: [[MUL:%.*]] = mul nuw nsw i32 [[CONV]], [[CONV]]<br>
+; CHECK-NEXT: [[MUL:%.*]] = mul nuw i16 [[CONV]], [[CONV]]<br>
; CHECK-NEXT: [[TINKY:%.*]] = load i16, i16* @glob, align 2<br>
-; CHECK-NEXT: [[TMP1:%.*]] = trunc i32 [[MUL]] to i16<br>
-; CHECK-NEXT: [[CONV14:%.*]] = and i16 [[TINKY]], [[TMP1]]<br>
+; CHECK-NEXT: [[CONV14:%.*]] = and i16 [[TINKY]], [[MUL]]<br>
; CHECK-NEXT: store i16 [[CONV14]], i16* @glob, align 2<br>
; CHECK-NEXT: ret void<br>
;<br>
<br>
<br>
______________________________<wbr>_________________<br>
llvm-commits mailing list<br>
<a href="mailto:llvm-commits@lists.llvm.org" target="_blank">llvm-commits@lists.llvm.org</a><br>
<a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits" rel="noreferrer" target="_blank">http://lists.llvm.org/cgi-bin/<wbr>mailman/listinfo/llvm-commits</a><br>
</blockquote></div></div></div></div>
</blockquote></div><br></div>
</div></div></blockquote></div><br></div>